I think we'll naturally develop intuition and infrastructure for fault-tolerant ways to implement LLMs in non-verifiable domains, i.e. requiring multiple, independent agents agree on a direction from multiple angles before taking an action.

But we're definitely very early on this trend.